1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* A network driver using virtio.
4 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
7 #include <linux/netdevice.h>
8 #include <linux/etherdevice.h>
9 #include <linux/ethtool.h>
10 #include <linux/module.h>
11 #include <linux/virtio.h>
12 #include <linux/virtio_net.h>
13 #include <linux/bpf.h>
14 #include <linux/bpf_trace.h>
15 #include <linux/scatterlist.h>
16 #include <linux/if_vlan.h>
17 #include <linux/slab.h>
18 #include <linux/cpu.h>
19 #include <linux/average.h>
20 #include <linux/filter.h>
21 #include <linux/kernel.h>
22 #include <net/route.h>
24 #include <net/net_failover.h>
26 static int napi_weight
= NAPI_POLL_WEIGHT
;
27 module_param(napi_weight
, int, 0444);
29 static bool csum
= true, gso
= true, napi_tx
= true;
30 module_param(csum
, bool, 0444);
31 module_param(gso
, bool, 0444);
32 module_param(napi_tx
, bool, 0644);
34 /* FIXME: MTU in config. */
35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
36 #define GOOD_COPY_LEN 128
38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
41 #define VIRTIO_XDP_HEADROOM 256
43 /* Separating two types of XDP xmit */
44 #define VIRTIO_XDP_TX BIT(0)
45 #define VIRTIO_XDP_REDIR BIT(1)
47 #define VIRTIO_XDP_FLAG BIT(0)
49 /* RX packet size EWMA. The average packet size is used to determine the packet
50 * buffer size when refilling RX rings. As the entire RX ring may be refilled
51 * at once, the weight is chosen so that the EWMA will be insensitive to short-
52 * term, transient changes in packet size.
54 DECLARE_EWMA(pkt_len
, 0, 64)
56 #define VIRTNET_DRIVER_VERSION "1.0.0"
58 static const unsigned long guest_offloads
[] = {
59 VIRTIO_NET_F_GUEST_TSO4
,
60 VIRTIO_NET_F_GUEST_TSO6
,
61 VIRTIO_NET_F_GUEST_ECN
,
62 VIRTIO_NET_F_GUEST_UFO
,
63 VIRTIO_NET_F_GUEST_CSUM
66 struct virtnet_stat_desc
{
67 char desc
[ETH_GSTRING_LEN
];
71 struct virtnet_sq_stats
{
72 struct u64_stats_sync syncp
;
80 struct virtnet_rq_stats
{
81 struct u64_stats_sync syncp
;
92 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
93 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
95 static const struct virtnet_stat_desc virtnet_sq_stats_desc
[] = {
96 { "packets", VIRTNET_SQ_STAT(packets
) },
97 { "bytes", VIRTNET_SQ_STAT(bytes
) },
98 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx
) },
99 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops
) },
100 { "kicks", VIRTNET_SQ_STAT(kicks
) },
103 static const struct virtnet_stat_desc virtnet_rq_stats_desc
[] = {
104 { "packets", VIRTNET_RQ_STAT(packets
) },
105 { "bytes", VIRTNET_RQ_STAT(bytes
) },
106 { "drops", VIRTNET_RQ_STAT(drops
) },
107 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets
) },
108 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx
) },
109 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects
) },
110 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops
) },
111 { "kicks", VIRTNET_RQ_STAT(kicks
) },
114 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
115 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
117 /* Internal representation of a send virtqueue */
119 /* Virtqueue associated with this send _queue */
120 struct virtqueue
*vq
;
122 /* TX: fragments + linear part + virtio header */
123 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
125 /* Name of the send queue: output.$index */
128 struct virtnet_sq_stats stats
;
130 struct napi_struct napi
;
133 /* Internal representation of a receive virtqueue */
134 struct receive_queue
{
135 /* Virtqueue associated with this receive_queue */
136 struct virtqueue
*vq
;
138 struct napi_struct napi
;
140 struct bpf_prog __rcu
*xdp_prog
;
142 struct virtnet_rq_stats stats
;
144 /* Chain pages by the private ptr. */
147 /* Average packet length for mergeable receive buffers. */
148 struct ewma_pkt_len mrg_avg_pkt_len
;
150 /* Page frag for packet buffer allocation. */
151 struct page_frag alloc_frag
;
153 /* RX: fragments + linear part + virtio header */
154 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
156 /* Min single buffer size for mergeable buffers case. */
157 unsigned int min_buf_len
;
159 /* Name of this receive queue: input.$index */
162 struct xdp_rxq_info xdp_rxq
;
165 /* Control VQ buffers: protected by the rtnl lock */
167 struct virtio_net_ctrl_hdr hdr
;
168 virtio_net_ctrl_ack status
;
169 struct virtio_net_ctrl_mq mq
;
176 struct virtnet_info
{
177 struct virtio_device
*vdev
;
178 struct virtqueue
*cvq
;
179 struct net_device
*dev
;
180 struct send_queue
*sq
;
181 struct receive_queue
*rq
;
184 /* Max # of queue pairs supported by the device */
187 /* # of queue pairs currently used by the driver */
188 u16 curr_queue_pairs
;
190 /* # of XDP queue pairs currently used by the driver */
193 /* I like... big packets and I cannot lie! */
196 /* Host will merge rx buffers for big packets (shake it! shake it!) */
197 bool mergeable_rx_bufs
;
199 /* Has control virtqueue */
202 /* Host can handle any s/g split between our header and packet data */
205 /* Packet virtio header size */
208 /* Work struct for refilling if we run low on memory. */
209 struct delayed_work refill
;
211 /* Work struct for config space updates */
212 struct work_struct config_work
;
214 /* Does the affinity hint is set for virtqueues? */
215 bool affinity_hint_set
;
217 /* CPU hotplug instances for online & dead */
218 struct hlist_node node
;
219 struct hlist_node node_dead
;
221 struct control_buf
*ctrl
;
223 /* Ethtool settings */
227 unsigned long guest_offloads
;
228 unsigned long guest_offloads_capable
;
230 /* failover when STANDBY feature enabled */
231 struct failover
*failover
;
234 struct padded_vnet_hdr
{
235 struct virtio_net_hdr_mrg_rxbuf hdr
;
237 * hdr is in a separate sg buffer, and data sg buffer shares same page
238 * with this header sg. This padding makes next sg 16 byte aligned
244 static bool is_xdp_frame(void *ptr
)
246 return (unsigned long)ptr
& VIRTIO_XDP_FLAG
;
249 static void *xdp_to_ptr(struct xdp_frame
*ptr
)
251 return (void *)((unsigned long)ptr
| VIRTIO_XDP_FLAG
);
254 static struct xdp_frame
*ptr_to_xdp(void *ptr
)
256 return (struct xdp_frame
*)((unsigned long)ptr
& ~VIRTIO_XDP_FLAG
);
259 /* Converting between virtqueue no. and kernel tx/rx queue no.
260 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
262 static int vq2txq(struct virtqueue
*vq
)
264 return (vq
->index
- 1) / 2;
267 static int txq2vq(int txq
)
272 static int vq2rxq(struct virtqueue
*vq
)
274 return vq
->index
/ 2;
277 static int rxq2vq(int rxq
)
282 static inline struct virtio_net_hdr_mrg_rxbuf
*skb_vnet_hdr(struct sk_buff
*skb
)
284 return (struct virtio_net_hdr_mrg_rxbuf
*)skb
->cb
;
288 * private is used to chain pages for big packets, put the whole
289 * most recent used list in the beginning for reuse
291 static void give_pages(struct receive_queue
*rq
, struct page
*page
)
295 /* Find end of list, sew whole thing into vi->rq.pages. */
296 for (end
= page
; end
->private; end
= (struct page
*)end
->private);
297 end
->private = (unsigned long)rq
->pages
;
301 static struct page
*get_a_page(struct receive_queue
*rq
, gfp_t gfp_mask
)
303 struct page
*p
= rq
->pages
;
306 rq
->pages
= (struct page
*)p
->private;
307 /* clear private here, it is used to chain pages */
310 p
= alloc_page(gfp_mask
);
314 static void virtqueue_napi_schedule(struct napi_struct
*napi
,
315 struct virtqueue
*vq
)
317 if (napi_schedule_prep(napi
)) {
318 virtqueue_disable_cb(vq
);
319 __napi_schedule(napi
);
323 static void virtqueue_napi_complete(struct napi_struct
*napi
,
324 struct virtqueue
*vq
, int processed
)
328 opaque
= virtqueue_enable_cb_prepare(vq
);
329 if (napi_complete_done(napi
, processed
)) {
330 if (unlikely(virtqueue_poll(vq
, opaque
)))
331 virtqueue_napi_schedule(napi
, vq
);
333 virtqueue_disable_cb(vq
);
337 static void skb_xmit_done(struct virtqueue
*vq
)
339 struct virtnet_info
*vi
= vq
->vdev
->priv
;
340 struct napi_struct
*napi
= &vi
->sq
[vq2txq(vq
)].napi
;
342 /* Suppress further interrupts. */
343 virtqueue_disable_cb(vq
);
346 virtqueue_napi_schedule(napi
, vq
);
348 /* We were probably waiting for more output buffers. */
349 netif_wake_subqueue(vi
->dev
, vq2txq(vq
));
352 #define MRG_CTX_HEADER_SHIFT 22
353 static void *mergeable_len_to_ctx(unsigned int truesize
,
354 unsigned int headroom
)
356 return (void *)(unsigned long)((headroom
<< MRG_CTX_HEADER_SHIFT
) | truesize
);
359 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx
)
361 return (unsigned long)mrg_ctx
>> MRG_CTX_HEADER_SHIFT
;
364 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx
)
366 return (unsigned long)mrg_ctx
& ((1 << MRG_CTX_HEADER_SHIFT
) - 1);
369 /* Called from bottom half context */
370 static struct sk_buff
*page_to_skb(struct virtnet_info
*vi
,
371 struct receive_queue
*rq
,
372 struct page
*page
, unsigned int offset
,
373 unsigned int len
, unsigned int truesize
,
374 bool hdr_valid
, unsigned int metasize
)
377 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
378 unsigned int copy
, hdr_len
, hdr_padded_len
;
381 p
= page_address(page
) + offset
;
383 /* copy small packet so we can reuse these pages for small data */
384 skb
= napi_alloc_skb(&rq
->napi
, GOOD_COPY_LEN
);
388 hdr
= skb_vnet_hdr(skb
);
390 hdr_len
= vi
->hdr_len
;
391 if (vi
->mergeable_rx_bufs
)
392 hdr_padded_len
= sizeof(*hdr
);
394 hdr_padded_len
= sizeof(struct padded_vnet_hdr
);
396 /* hdr_valid means no XDP, so we can copy the vnet header */
398 memcpy(hdr
, p
, hdr_len
);
401 offset
+= hdr_padded_len
;
405 if (copy
> skb_tailroom(skb
))
406 copy
= skb_tailroom(skb
);
407 skb_put_data(skb
, p
, copy
);
410 __skb_pull(skb
, metasize
);
411 skb_metadata_set(skb
, metasize
);
417 if (vi
->mergeable_rx_bufs
) {
419 skb_add_rx_frag(skb
, 0, page
, offset
, len
, truesize
);
426 * Verify that we can indeed put this data into a skb.
427 * This is here to handle cases when the device erroneously
428 * tries to receive more than is possible. This is usually
429 * the case of a broken device.
431 if (unlikely(len
> MAX_SKB_FRAGS
* PAGE_SIZE
)) {
432 net_dbg_ratelimited("%s: too much data\n", skb
->dev
->name
);
436 BUG_ON(offset
>= PAGE_SIZE
);
438 unsigned int frag_size
= min((unsigned)PAGE_SIZE
- offset
, len
);
439 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
, offset
,
440 frag_size
, truesize
);
442 page
= (struct page
*)page
->private;
447 give_pages(rq
, page
);
452 static int __virtnet_xdp_xmit_one(struct virtnet_info
*vi
,
453 struct send_queue
*sq
,
454 struct xdp_frame
*xdpf
)
456 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
459 if (unlikely(xdpf
->headroom
< vi
->hdr_len
))
462 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
463 xdpf
->data
-= vi
->hdr_len
;
464 /* Zero header and leave csum up to XDP layers */
466 memset(hdr
, 0, vi
->hdr_len
);
467 xdpf
->len
+= vi
->hdr_len
;
469 sg_init_one(sq
->sg
, xdpf
->data
, xdpf
->len
);
471 err
= virtqueue_add_outbuf(sq
->vq
, sq
->sg
, 1, xdp_to_ptr(xdpf
),
474 return -ENOSPC
; /* Caller handle free/refcnt */
479 static struct send_queue
*virtnet_xdp_sq(struct virtnet_info
*vi
)
483 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
487 static int virtnet_xdp_xmit(struct net_device
*dev
,
488 int n
, struct xdp_frame
**frames
, u32 flags
)
490 struct virtnet_info
*vi
= netdev_priv(dev
);
491 struct receive_queue
*rq
= vi
->rq
;
492 struct bpf_prog
*xdp_prog
;
493 struct send_queue
*sq
;
503 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
504 * indicate XDP resources have been successfully allocated.
506 xdp_prog
= rcu_access_pointer(rq
->xdp_prog
);
510 sq
= virtnet_xdp_sq(vi
);
512 if (unlikely(flags
& ~XDP_XMIT_FLAGS_MASK
)) {
518 /* Free up any pending old buffers before queueing new ones. */
519 while ((ptr
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
520 if (likely(is_xdp_frame(ptr
))) {
521 struct xdp_frame
*frame
= ptr_to_xdp(ptr
);
524 xdp_return_frame(frame
);
526 struct sk_buff
*skb
= ptr
;
529 napi_consume_skb(skb
, false);
534 for (i
= 0; i
< n
; i
++) {
535 struct xdp_frame
*xdpf
= frames
[i
];
537 err
= __virtnet_xdp_xmit_one(vi
, sq
, xdpf
);
539 xdp_return_frame_rx_napi(xdpf
);
545 if (flags
& XDP_XMIT_FLUSH
) {
546 if (virtqueue_kick_prepare(sq
->vq
) && virtqueue_notify(sq
->vq
))
550 u64_stats_update_begin(&sq
->stats
.syncp
);
551 sq
->stats
.bytes
+= bytes
;
552 sq
->stats
.packets
+= packets
;
553 sq
->stats
.xdp_tx
+= n
;
554 sq
->stats
.xdp_tx_drops
+= drops
;
555 sq
->stats
.kicks
+= kicks
;
556 u64_stats_update_end(&sq
->stats
.syncp
);
561 static unsigned int virtnet_get_headroom(struct virtnet_info
*vi
)
563 return vi
->xdp_queue_pairs
? VIRTIO_XDP_HEADROOM
: 0;
566 /* We copy the packet for XDP in the following cases:
568 * 1) Packet is scattered across multiple rx buffers.
569 * 2) Headroom space is insufficient.
571 * This is inefficient but it's a temporary condition that
572 * we hit right after XDP is enabled and until queue is refilled
573 * with large buffers with sufficient headroom - so it should affect
574 * at most queue size packets.
575 * Afterwards, the conditions to enable
576 * XDP should preclude the underlying device from sending packets
577 * across multiple buffers (num_buf > 1), and we make sure buffers
578 * have enough headroom.
580 static struct page
*xdp_linearize_page(struct receive_queue
*rq
,
587 struct page
*page
= alloc_page(GFP_ATOMIC
);
592 memcpy(page_address(page
) + page_off
, page_address(p
) + offset
, *len
);
596 int tailroom
= SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
601 buf
= virtqueue_get_buf(rq
->vq
, &buflen
);
605 p
= virt_to_head_page(buf
);
606 off
= buf
- page_address(p
);
608 /* guard against a misconfigured or uncooperative backend that
609 * is sending packet larger than the MTU.
611 if ((page_off
+ buflen
+ tailroom
) > PAGE_SIZE
) {
616 memcpy(page_address(page
) + page_off
,
617 page_address(p
) + off
, buflen
);
622 /* Headroom does not contribute to packet length */
623 *len
= page_off
- VIRTIO_XDP_HEADROOM
;
626 __free_pages(page
, 0);
630 static struct sk_buff
*receive_small(struct net_device
*dev
,
631 struct virtnet_info
*vi
,
632 struct receive_queue
*rq
,
633 void *buf
, void *ctx
,
635 unsigned int *xdp_xmit
,
636 struct virtnet_rq_stats
*stats
)
639 struct bpf_prog
*xdp_prog
;
640 unsigned int xdp_headroom
= (unsigned long)ctx
;
641 unsigned int header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
642 unsigned int headroom
= vi
->hdr_len
+ header_offset
;
643 unsigned int buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
644 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
645 struct page
*page
= virt_to_head_page(buf
);
646 unsigned int delta
= 0;
647 struct page
*xdp_page
;
649 unsigned int metasize
= 0;
655 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
657 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
+ header_offset
;
658 struct xdp_frame
*xdpf
;
663 if (unlikely(hdr
->hdr
.gso_type
))
666 if (unlikely(xdp_headroom
< virtnet_get_headroom(vi
))) {
667 int offset
= buf
- page_address(page
) + header_offset
;
668 unsigned int tlen
= len
+ vi
->hdr_len
;
671 xdp_headroom
= virtnet_get_headroom(vi
);
672 header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
673 headroom
= vi
->hdr_len
+ header_offset
;
674 buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
675 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
676 xdp_page
= xdp_linearize_page(rq
, &num_buf
, page
,
677 offset
, header_offset
,
682 buf
= page_address(xdp_page
);
687 xdp
.data_hard_start
= buf
+ VIRTNET_RX_PAD
+ vi
->hdr_len
;
688 xdp
.data
= xdp
.data_hard_start
+ xdp_headroom
;
689 xdp
.data_end
= xdp
.data
+ len
;
690 xdp
.data_meta
= xdp
.data
;
691 xdp
.rxq
= &rq
->xdp_rxq
;
692 orig_data
= xdp
.data
;
693 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
694 stats
->xdp_packets
++;
698 /* Recalculate length in case bpf program changed it */
699 delta
= orig_data
- xdp
.data
;
700 len
= xdp
.data_end
- xdp
.data
;
701 metasize
= xdp
.data
- xdp
.data_meta
;
705 xdpf
= convert_to_xdp_frame(&xdp
);
708 err
= virtnet_xdp_xmit(dev
, 1, &xdpf
, 0);
709 if (unlikely(err
< 0)) {
710 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
713 *xdp_xmit
|= VIRTIO_XDP_TX
;
717 stats
->xdp_redirects
++;
718 err
= xdp_do_redirect(dev
, &xdp
, xdp_prog
);
721 *xdp_xmit
|= VIRTIO_XDP_REDIR
;
725 bpf_warn_invalid_xdp_action(act
);
728 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
735 skb
= build_skb(buf
, buflen
);
740 skb_reserve(skb
, headroom
- delta
);
743 buf
+= header_offset
;
744 memcpy(skb_vnet_hdr(skb
), buf
, vi
->hdr_len
);
745 } /* keep zeroed vnet hdr since XDP is loaded */
748 skb_metadata_set(skb
, metasize
);
762 static struct sk_buff
*receive_big(struct net_device
*dev
,
763 struct virtnet_info
*vi
,
764 struct receive_queue
*rq
,
767 struct virtnet_rq_stats
*stats
)
769 struct page
*page
= buf
;
770 struct sk_buff
*skb
=
771 page_to_skb(vi
, rq
, page
, 0, len
, PAGE_SIZE
, true, 0);
773 stats
->bytes
+= len
- vi
->hdr_len
;
781 give_pages(rq
, page
);
785 static struct sk_buff
*receive_mergeable(struct net_device
*dev
,
786 struct virtnet_info
*vi
,
787 struct receive_queue
*rq
,
791 unsigned int *xdp_xmit
,
792 struct virtnet_rq_stats
*stats
)
794 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
795 u16 num_buf
= virtio16_to_cpu(vi
->vdev
, hdr
->num_buffers
);
796 struct page
*page
= virt_to_head_page(buf
);
797 int offset
= buf
- page_address(page
);
798 struct sk_buff
*head_skb
, *curr_skb
;
799 struct bpf_prog
*xdp_prog
;
800 unsigned int truesize
;
801 unsigned int headroom
= mergeable_ctx_to_headroom(ctx
);
803 unsigned int metasize
= 0;
806 stats
->bytes
+= len
- vi
->hdr_len
;
809 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
811 struct xdp_frame
*xdpf
;
812 struct page
*xdp_page
;
817 /* Transient failure which in theory could occur if
818 * in-flight packets from before XDP was enabled reach
819 * the receive path after XDP is loaded.
821 if (unlikely(hdr
->hdr
.gso_type
))
824 /* This happens when rx buffer size is underestimated
825 * or headroom is not enough because of the buffer
826 * was refilled before XDP is set. This should only
827 * happen for the first several packets, so we don't
828 * care much about its performance.
830 if (unlikely(num_buf
> 1 ||
831 headroom
< virtnet_get_headroom(vi
))) {
832 /* linearize data for XDP */
833 xdp_page
= xdp_linearize_page(rq
, &num_buf
,
839 offset
= VIRTIO_XDP_HEADROOM
;
844 /* Allow consuming headroom but reserve enough space to push
845 * the descriptor on if we get an XDP_TX return code.
847 data
= page_address(xdp_page
) + offset
;
848 xdp
.data_hard_start
= data
- VIRTIO_XDP_HEADROOM
+ vi
->hdr_len
;
849 xdp
.data
= data
+ vi
->hdr_len
;
850 xdp
.data_end
= xdp
.data
+ (len
- vi
->hdr_len
);
851 xdp
.data_meta
= xdp
.data
;
852 xdp
.rxq
= &rq
->xdp_rxq
;
854 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
855 stats
->xdp_packets
++;
859 metasize
= xdp
.data
- xdp
.data_meta
;
861 /* recalculate offset to account for any header
862 * adjustments and minus the metasize to copy the
863 * metadata in page_to_skb(). Note other cases do not
864 * build an skb and avoid using offset
866 offset
= xdp
.data
- page_address(xdp_page
) -
867 vi
->hdr_len
- metasize
;
869 /* recalculate len if xdp.data, xdp.data_end or
870 * xdp.data_meta were adjusted
872 len
= xdp
.data_end
- xdp
.data
+ vi
->hdr_len
+ metasize
;
873 /* We can only create skb based on xdp_page. */
874 if (unlikely(xdp_page
!= page
)) {
877 head_skb
= page_to_skb(vi
, rq
, xdp_page
, offset
,
878 len
, PAGE_SIZE
, false,
885 xdpf
= convert_to_xdp_frame(&xdp
);
888 err
= virtnet_xdp_xmit(dev
, 1, &xdpf
, 0);
889 if (unlikely(err
< 0)) {
890 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
891 if (unlikely(xdp_page
!= page
))
895 *xdp_xmit
|= VIRTIO_XDP_TX
;
896 if (unlikely(xdp_page
!= page
))
901 stats
->xdp_redirects
++;
902 err
= xdp_do_redirect(dev
, &xdp
, xdp_prog
);
904 if (unlikely(xdp_page
!= page
))
908 *xdp_xmit
|= VIRTIO_XDP_REDIR
;
909 if (unlikely(xdp_page
!= page
))
914 bpf_warn_invalid_xdp_action(act
);
917 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
920 if (unlikely(xdp_page
!= page
))
921 __free_pages(xdp_page
, 0);
927 truesize
= mergeable_ctx_to_truesize(ctx
);
928 if (unlikely(len
> truesize
)) {
929 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
930 dev
->name
, len
, (unsigned long)ctx
);
931 dev
->stats
.rx_length_errors
++;
935 head_skb
= page_to_skb(vi
, rq
, page
, offset
, len
, truesize
, !xdp_prog
,
939 if (unlikely(!curr_skb
))
944 buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
);
945 if (unlikely(!buf
)) {
946 pr_debug("%s: rx error: %d buffers out of %d missing\n",
948 virtio16_to_cpu(vi
->vdev
,
950 dev
->stats
.rx_length_errors
++;
955 page
= virt_to_head_page(buf
);
957 truesize
= mergeable_ctx_to_truesize(ctx
);
958 if (unlikely(len
> truesize
)) {
959 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
960 dev
->name
, len
, (unsigned long)ctx
);
961 dev
->stats
.rx_length_errors
++;
965 num_skb_frags
= skb_shinfo(curr_skb
)->nr_frags
;
966 if (unlikely(num_skb_frags
== MAX_SKB_FRAGS
)) {
967 struct sk_buff
*nskb
= alloc_skb(0, GFP_ATOMIC
);
971 if (curr_skb
== head_skb
)
972 skb_shinfo(curr_skb
)->frag_list
= nskb
;
974 curr_skb
->next
= nskb
;
976 head_skb
->truesize
+= nskb
->truesize
;
979 if (curr_skb
!= head_skb
) {
980 head_skb
->data_len
+= len
;
981 head_skb
->len
+= len
;
982 head_skb
->truesize
+= truesize
;
984 offset
= buf
- page_address(page
);
985 if (skb_can_coalesce(curr_skb
, num_skb_frags
, page
, offset
)) {
987 skb_coalesce_rx_frag(curr_skb
, num_skb_frags
- 1,
990 skb_add_rx_frag(curr_skb
, num_skb_frags
, page
,
991 offset
, len
, truesize
);
995 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, head_skb
->len
);
1003 while (num_buf
-- > 1) {
1004 buf
= virtqueue_get_buf(rq
->vq
, &len
);
1005 if (unlikely(!buf
)) {
1006 pr_debug("%s: rx error: %d buffers missing\n",
1007 dev
->name
, num_buf
);
1008 dev
->stats
.rx_length_errors
++;
1011 stats
->bytes
+= len
;
1012 page
= virt_to_head_page(buf
);
1017 dev_kfree_skb(head_skb
);
1022 static void receive_buf(struct virtnet_info
*vi
, struct receive_queue
*rq
,
1023 void *buf
, unsigned int len
, void **ctx
,
1024 unsigned int *xdp_xmit
,
1025 struct virtnet_rq_stats
*stats
)
1027 struct net_device
*dev
= vi
->dev
;
1028 struct sk_buff
*skb
;
1029 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1031 if (unlikely(len
< vi
->hdr_len
+ ETH_HLEN
)) {
1032 pr_debug("%s: short packet %i\n", dev
->name
, len
);
1033 dev
->stats
.rx_length_errors
++;
1034 if (vi
->mergeable_rx_bufs
) {
1035 put_page(virt_to_head_page(buf
));
1036 } else if (vi
->big_packets
) {
1037 give_pages(rq
, buf
);
1039 put_page(virt_to_head_page(buf
));
1044 if (vi
->mergeable_rx_bufs
)
1045 skb
= receive_mergeable(dev
, vi
, rq
, buf
, ctx
, len
, xdp_xmit
,
1047 else if (vi
->big_packets
)
1048 skb
= receive_big(dev
, vi
, rq
, buf
, len
, stats
);
1050 skb
= receive_small(dev
, vi
, rq
, buf
, ctx
, len
, xdp_xmit
, stats
);
1055 hdr
= skb_vnet_hdr(skb
);
1057 if (hdr
->hdr
.flags
& VIRTIO_NET_HDR_F_DATA_VALID
)
1058 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
1060 if (virtio_net_hdr_to_skb(skb
, &hdr
->hdr
,
1061 virtio_is_little_endian(vi
->vdev
))) {
1062 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1063 dev
->name
, hdr
->hdr
.gso_type
,
1068 skb_record_rx_queue(skb
, vq2rxq(rq
->vq
));
1069 skb
->protocol
= eth_type_trans(skb
, dev
);
1070 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1071 ntohs(skb
->protocol
), skb
->len
, skb
->pkt_type
);
1073 napi_gro_receive(&rq
->napi
, skb
);
1077 dev
->stats
.rx_frame_errors
++;
1081 /* Unlike mergeable buffers, all buffers are allocated to the
1082 * same size, except for the headroom. For this reason we do
1083 * not need to use mergeable_len_to_ctx here - it is enough
1084 * to store the headroom as the context ignoring the truesize.
1086 static int add_recvbuf_small(struct virtnet_info
*vi
, struct receive_queue
*rq
,
1089 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
1091 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
1092 void *ctx
= (void *)(unsigned long)xdp_headroom
;
1093 int len
= vi
->hdr_len
+ VIRTNET_RX_PAD
+ GOOD_PACKET_LEN
+ xdp_headroom
;
1096 len
= SKB_DATA_ALIGN(len
) +
1097 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
1098 if (unlikely(!skb_page_frag_refill(len
, alloc_frag
, gfp
)))
1101 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
1102 get_page(alloc_frag
->page
);
1103 alloc_frag
->offset
+= len
;
1104 sg_init_one(rq
->sg
, buf
+ VIRTNET_RX_PAD
+ xdp_headroom
,
1105 vi
->hdr_len
+ GOOD_PACKET_LEN
);
1106 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
1108 put_page(virt_to_head_page(buf
));
1112 static int add_recvbuf_big(struct virtnet_info
*vi
, struct receive_queue
*rq
,
1115 struct page
*first
, *list
= NULL
;
1119 sg_init_table(rq
->sg
, MAX_SKB_FRAGS
+ 2);
1121 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1122 for (i
= MAX_SKB_FRAGS
+ 1; i
> 1; --i
) {
1123 first
= get_a_page(rq
, gfp
);
1126 give_pages(rq
, list
);
1129 sg_set_buf(&rq
->sg
[i
], page_address(first
), PAGE_SIZE
);
1131 /* chain new page in list head to match sg */
1132 first
->private = (unsigned long)list
;
1136 first
= get_a_page(rq
, gfp
);
1138 give_pages(rq
, list
);
1141 p
= page_address(first
);
1143 /* rq->sg[0], rq->sg[1] share the same page */
1144 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1145 sg_set_buf(&rq
->sg
[0], p
, vi
->hdr_len
);
1147 /* rq->sg[1] for data packet, from offset */
1148 offset
= sizeof(struct padded_vnet_hdr
);
1149 sg_set_buf(&rq
->sg
[1], p
+ offset
, PAGE_SIZE
- offset
);
1151 /* chain first in list head */
1152 first
->private = (unsigned long)list
;
1153 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, MAX_SKB_FRAGS
+ 2,
1156 give_pages(rq
, first
);
1161 static unsigned int get_mergeable_buf_len(struct receive_queue
*rq
,
1162 struct ewma_pkt_len
*avg_pkt_len
,
1165 const size_t hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
1169 return PAGE_SIZE
- room
;
1171 len
= hdr_len
+ clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len
),
1172 rq
->min_buf_len
, PAGE_SIZE
- hdr_len
);
1174 return ALIGN(len
, L1_CACHE_BYTES
);
1177 static int add_recvbuf_mergeable(struct virtnet_info
*vi
,
1178 struct receive_queue
*rq
, gfp_t gfp
)
1180 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
1181 unsigned int headroom
= virtnet_get_headroom(vi
);
1182 unsigned int tailroom
= headroom
? sizeof(struct skb_shared_info
) : 0;
1183 unsigned int room
= SKB_DATA_ALIGN(headroom
+ tailroom
);
1187 unsigned int len
, hole
;
1189 /* Extra tailroom is needed to satisfy XDP's assumption. This
1190 * means rx frags coalescing won't work, but consider we've
1191 * disabled GSO for XDP, it won't be a big issue.
1193 len
= get_mergeable_buf_len(rq
, &rq
->mrg_avg_pkt_len
, room
);
1194 if (unlikely(!skb_page_frag_refill(len
+ room
, alloc_frag
, gfp
)))
1197 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
1198 buf
+= headroom
; /* advance address leaving hole at front of pkt */
1199 get_page(alloc_frag
->page
);
1200 alloc_frag
->offset
+= len
+ room
;
1201 hole
= alloc_frag
->size
- alloc_frag
->offset
;
1202 if (hole
< len
+ room
) {
1203 /* To avoid internal fragmentation, if there is very likely not
1204 * enough space for another buffer, add the remaining space to
1205 * the current buffer.
1208 alloc_frag
->offset
+= hole
;
1211 sg_init_one(rq
->sg
, buf
, len
);
1212 ctx
= mergeable_len_to_ctx(len
, headroom
);
1213 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
1215 put_page(virt_to_head_page(buf
));
1221 * Returns false if we couldn't fill entirely (OOM).
1223 * Normally run in the receive path, but can also be run from ndo_open
1224 * before we're receiving packets, or from refill_work which is
1225 * careful to disable receiving (using napi_disable).
1227 static bool try_fill_recv(struct virtnet_info
*vi
, struct receive_queue
*rq
,
1234 if (vi
->mergeable_rx_bufs
)
1235 err
= add_recvbuf_mergeable(vi
, rq
, gfp
);
1236 else if (vi
->big_packets
)
1237 err
= add_recvbuf_big(vi
, rq
, gfp
);
1239 err
= add_recvbuf_small(vi
, rq
, gfp
);
1241 oom
= err
== -ENOMEM
;
1244 } while (rq
->vq
->num_free
);
1245 if (virtqueue_kick_prepare(rq
->vq
) && virtqueue_notify(rq
->vq
)) {
1246 u64_stats_update_begin(&rq
->stats
.syncp
);
1248 u64_stats_update_end(&rq
->stats
.syncp
);
1254 static void skb_recv_done(struct virtqueue
*rvq
)
1256 struct virtnet_info
*vi
= rvq
->vdev
->priv
;
1257 struct receive_queue
*rq
= &vi
->rq
[vq2rxq(rvq
)];
1259 virtqueue_napi_schedule(&rq
->napi
, rvq
);
1262 static void virtnet_napi_enable(struct virtqueue
*vq
, struct napi_struct
*napi
)
1266 /* If all buffers were filled by other side before we napi_enabled, we
1267 * won't get another interrupt, so process any outstanding packets now.
1268 * Call local_bh_enable after to trigger softIRQ processing.
1271 virtqueue_napi_schedule(napi
, vq
);
1275 static void virtnet_napi_tx_enable(struct virtnet_info
*vi
,
1276 struct virtqueue
*vq
,
1277 struct napi_struct
*napi
)
1282 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1283 * enable the feature if this is likely affine with the transmit path.
1285 if (!vi
->affinity_hint_set
) {
1290 return virtnet_napi_enable(vq
, napi
);
1293 static void virtnet_napi_tx_disable(struct napi_struct
*napi
)
1299 static void refill_work(struct work_struct
*work
)
1301 struct virtnet_info
*vi
=
1302 container_of(work
, struct virtnet_info
, refill
.work
);
1306 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
1307 struct receive_queue
*rq
= &vi
->rq
[i
];
1309 napi_disable(&rq
->napi
);
1310 still_empty
= !try_fill_recv(vi
, rq
, GFP_KERNEL
);
1311 virtnet_napi_enable(rq
->vq
, &rq
->napi
);
1313 /* In theory, this can happen: if we don't get any buffers in
1314 * we will *never* try to fill again.
1317 schedule_delayed_work(&vi
->refill
, HZ
/2);
1321 static int virtnet_receive(struct receive_queue
*rq
, int budget
,
1322 unsigned int *xdp_xmit
)
1324 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1325 struct virtnet_rq_stats stats
= {};
1330 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
1333 while (stats
.packets
< budget
&&
1334 (buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
))) {
1335 receive_buf(vi
, rq
, buf
, len
, ctx
, xdp_xmit
, &stats
);
1339 while (stats
.packets
< budget
&&
1340 (buf
= virtqueue_get_buf(rq
->vq
, &len
)) != NULL
) {
1341 receive_buf(vi
, rq
, buf
, len
, NULL
, xdp_xmit
, &stats
);
1346 if (rq
->vq
->num_free
> min((unsigned int)budget
, virtqueue_get_vring_size(rq
->vq
)) / 2) {
1347 if (!try_fill_recv(vi
, rq
, GFP_ATOMIC
))
1348 schedule_delayed_work(&vi
->refill
, 0);
1351 u64_stats_update_begin(&rq
->stats
.syncp
);
1352 for (i
= 0; i
< VIRTNET_RQ_STATS_LEN
; i
++) {
1353 size_t offset
= virtnet_rq_stats_desc
[i
].offset
;
1356 item
= (u64
*)((u8
*)&rq
->stats
+ offset
);
1357 *item
+= *(u64
*)((u8
*)&stats
+ offset
);
1359 u64_stats_update_end(&rq
->stats
.syncp
);
1361 return stats
.packets
;
1364 static void free_old_xmit_skbs(struct send_queue
*sq
, bool in_napi
)
1367 unsigned int packets
= 0;
1368 unsigned int bytes
= 0;
1371 while ((ptr
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
1372 if (likely(!is_xdp_frame(ptr
))) {
1373 struct sk_buff
*skb
= ptr
;
1375 pr_debug("Sent skb %p\n", skb
);
1378 napi_consume_skb(skb
, in_napi
);
1380 struct xdp_frame
*frame
= ptr_to_xdp(ptr
);
1382 bytes
+= frame
->len
;
1383 xdp_return_frame(frame
);
1388 /* Avoid overhead when no packets have been processed
1389 * happens when called speculatively from start_xmit.
1394 u64_stats_update_begin(&sq
->stats
.syncp
);
1395 sq
->stats
.bytes
+= bytes
;
1396 sq
->stats
.packets
+= packets
;
1397 u64_stats_update_end(&sq
->stats
.syncp
);
1400 static bool is_xdp_raw_buffer_queue(struct virtnet_info
*vi
, int q
)
1402 if (q
< (vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
))
1404 else if (q
< vi
->curr_queue_pairs
)
1410 static void virtnet_poll_cleantx(struct receive_queue
*rq
)
1412 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1413 unsigned int index
= vq2rxq(rq
->vq
);
1414 struct send_queue
*sq
= &vi
->sq
[index
];
1415 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, index
);
1417 if (!sq
->napi
.weight
|| is_xdp_raw_buffer_queue(vi
, index
))
1420 if (__netif_tx_trylock(txq
)) {
1421 free_old_xmit_skbs(sq
, true);
1422 __netif_tx_unlock(txq
);
1425 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1426 netif_tx_wake_queue(txq
);
1429 static int virtnet_poll(struct napi_struct
*napi
, int budget
)
1431 struct receive_queue
*rq
=
1432 container_of(napi
, struct receive_queue
, napi
);
1433 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1434 struct send_queue
*sq
;
1435 unsigned int received
;
1436 unsigned int xdp_xmit
= 0;
1438 virtnet_poll_cleantx(rq
);
1440 received
= virtnet_receive(rq
, budget
, &xdp_xmit
);
1442 /* Out of packets? */
1443 if (received
< budget
)
1444 virtqueue_napi_complete(napi
, rq
->vq
, received
);
1446 if (xdp_xmit
& VIRTIO_XDP_REDIR
)
1449 if (xdp_xmit
& VIRTIO_XDP_TX
) {
1450 sq
= virtnet_xdp_sq(vi
);
1451 if (virtqueue_kick_prepare(sq
->vq
) && virtqueue_notify(sq
->vq
)) {
1452 u64_stats_update_begin(&sq
->stats
.syncp
);
1454 u64_stats_update_end(&sq
->stats
.syncp
);
1461 static int virtnet_open(struct net_device
*dev
)
1463 struct virtnet_info
*vi
= netdev_priv(dev
);
1466 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1467 if (i
< vi
->curr_queue_pairs
)
1468 /* Make sure we have some buffers: if oom use wq. */
1469 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1470 schedule_delayed_work(&vi
->refill
, 0);
1472 err
= xdp_rxq_info_reg(&vi
->rq
[i
].xdp_rxq
, dev
, i
);
1476 err
= xdp_rxq_info_reg_mem_model(&vi
->rq
[i
].xdp_rxq
,
1477 MEM_TYPE_PAGE_SHARED
, NULL
);
1479 xdp_rxq_info_unreg(&vi
->rq
[i
].xdp_rxq
);
1483 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1484 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
, &vi
->sq
[i
].napi
);
1490 static int virtnet_poll_tx(struct napi_struct
*napi
, int budget
)
1492 struct send_queue
*sq
= container_of(napi
, struct send_queue
, napi
);
1493 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1494 unsigned int index
= vq2txq(sq
->vq
);
1495 struct netdev_queue
*txq
;
1497 if (unlikely(is_xdp_raw_buffer_queue(vi
, index
))) {
1498 /* We don't need to enable cb for XDP */
1499 napi_complete_done(napi
, 0);
1503 txq
= netdev_get_tx_queue(vi
->dev
, index
);
1504 __netif_tx_lock(txq
, raw_smp_processor_id());
1505 free_old_xmit_skbs(sq
, true);
1506 __netif_tx_unlock(txq
);
1508 virtqueue_napi_complete(napi
, sq
->vq
, 0);
1510 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1511 netif_tx_wake_queue(txq
);
1516 static int xmit_skb(struct send_queue
*sq
, struct sk_buff
*skb
)
1518 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1519 const unsigned char *dest
= ((struct ethhdr
*)skb
->data
)->h_dest
;
1520 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1522 unsigned hdr_len
= vi
->hdr_len
;
1525 pr_debug("%s: xmit %p %pM\n", vi
->dev
->name
, skb
, dest
);
1527 can_push
= vi
->any_header_sg
&&
1528 !((unsigned long)skb
->data
& (__alignof__(*hdr
) - 1)) &&
1529 !skb_header_cloned(skb
) && skb_headroom(skb
) >= hdr_len
;
1530 /* Even if we can, don't push here yet as this would skew
1531 * csum_start offset below. */
1533 hdr
= (struct virtio_net_hdr_mrg_rxbuf
*)(skb
->data
- hdr_len
);
1535 hdr
= skb_vnet_hdr(skb
);
1537 if (virtio_net_hdr_from_skb(skb
, &hdr
->hdr
,
1538 virtio_is_little_endian(vi
->vdev
), false,
1542 if (vi
->mergeable_rx_bufs
)
1543 hdr
->num_buffers
= 0;
1545 sg_init_table(sq
->sg
, skb_shinfo(skb
)->nr_frags
+ (can_push
? 1 : 2));
1547 __skb_push(skb
, hdr_len
);
1548 num_sg
= skb_to_sgvec(skb
, sq
->sg
, 0, skb
->len
);
1549 if (unlikely(num_sg
< 0))
1551 /* Pull header back to avoid skew in tx bytes calculations. */
1552 __skb_pull(skb
, hdr_len
);
1554 sg_set_buf(sq
->sg
, hdr
, hdr_len
);
1555 num_sg
= skb_to_sgvec(skb
, sq
->sg
+ 1, 0, skb
->len
);
1556 if (unlikely(num_sg
< 0))
1560 return virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
, skb
, GFP_ATOMIC
);
1563 static netdev_tx_t
start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
1565 struct virtnet_info
*vi
= netdev_priv(dev
);
1566 int qnum
= skb_get_queue_mapping(skb
);
1567 struct send_queue
*sq
= &vi
->sq
[qnum
];
1569 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, qnum
);
1570 bool kick
= !netdev_xmit_more();
1571 bool use_napi
= sq
->napi
.weight
;
1573 /* Free up any pending old buffers before queueing new ones. */
1574 free_old_xmit_skbs(sq
, false);
1576 if (use_napi
&& kick
)
1577 virtqueue_enable_cb_delayed(sq
->vq
);
1579 /* timestamp packet in software */
1580 skb_tx_timestamp(skb
);
1582 /* Try to transmit */
1583 err
= xmit_skb(sq
, skb
);
1585 /* This should not happen! */
1586 if (unlikely(err
)) {
1587 dev
->stats
.tx_fifo_errors
++;
1588 if (net_ratelimit())
1590 "Unexpected TXQ (%d) queue failure: %d\n",
1592 dev
->stats
.tx_dropped
++;
1593 dev_kfree_skb_any(skb
);
1594 return NETDEV_TX_OK
;
1597 /* Don't wait up for transmitted skbs to be freed. */
1603 /* If running out of space, stop queue to avoid getting packets that we
1604 * are then unable to transmit.
1605 * An alternative would be to force queuing layer to requeue the skb by
1606 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1607 * returned in a normal path of operation: it means that driver is not
1608 * maintaining the TX queue stop/start state properly, and causes
1609 * the stack to do a non-trivial amount of useless work.
1610 * Since most packets only take 1 or 2 ring slots, stopping the queue
1611 * early means 16 slots are typically wasted.
1613 if (sq
->vq
->num_free
< 2+MAX_SKB_FRAGS
) {
1614 netif_stop_subqueue(dev
, qnum
);
1616 unlikely(!virtqueue_enable_cb_delayed(sq
->vq
))) {
1617 /* More just got used, free them then recheck. */
1618 free_old_xmit_skbs(sq
, false);
1619 if (sq
->vq
->num_free
>= 2+MAX_SKB_FRAGS
) {
1620 netif_start_subqueue(dev
, qnum
);
1621 virtqueue_disable_cb(sq
->vq
);
1626 if (kick
|| netif_xmit_stopped(txq
)) {
1627 if (virtqueue_kick_prepare(sq
->vq
) && virtqueue_notify(sq
->vq
)) {
1628 u64_stats_update_begin(&sq
->stats
.syncp
);
1630 u64_stats_update_end(&sq
->stats
.syncp
);
1634 return NETDEV_TX_OK
;
1638 * Send command via the control virtqueue and check status. Commands
1639 * supported by the hypervisor, as indicated by feature bits, should
1640 * never fail unless improperly formatted.
1642 static bool virtnet_send_command(struct virtnet_info
*vi
, u8
class, u8 cmd
,
1643 struct scatterlist
*out
)
1645 struct scatterlist
*sgs
[4], hdr
, stat
;
1646 unsigned out_num
= 0, tmp
;
1648 /* Caller should know better */
1649 BUG_ON(!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
));
1651 vi
->ctrl
->status
= ~0;
1652 vi
->ctrl
->hdr
.class = class;
1653 vi
->ctrl
->hdr
.cmd
= cmd
;
1655 sg_init_one(&hdr
, &vi
->ctrl
->hdr
, sizeof(vi
->ctrl
->hdr
));
1656 sgs
[out_num
++] = &hdr
;
1659 sgs
[out_num
++] = out
;
1661 /* Add return status. */
1662 sg_init_one(&stat
, &vi
->ctrl
->status
, sizeof(vi
->ctrl
->status
));
1663 sgs
[out_num
] = &stat
;
1665 BUG_ON(out_num
+ 1 > ARRAY_SIZE(sgs
));
1666 virtqueue_add_sgs(vi
->cvq
, sgs
, out_num
, 1, vi
, GFP_ATOMIC
);
1668 if (unlikely(!virtqueue_kick(vi
->cvq
)))
1669 return vi
->ctrl
->status
== VIRTIO_NET_OK
;
1671 /* Spin for a response, the kick causes an ioport write, trapping
1672 * into the hypervisor, so the request should be handled immediately.
1674 while (!virtqueue_get_buf(vi
->cvq
, &tmp
) &&
1675 !virtqueue_is_broken(vi
->cvq
))
1678 return vi
->ctrl
->status
== VIRTIO_NET_OK
;
1681 static int virtnet_set_mac_address(struct net_device
*dev
, void *p
)
1683 struct virtnet_info
*vi
= netdev_priv(dev
);
1684 struct virtio_device
*vdev
= vi
->vdev
;
1686 struct sockaddr
*addr
;
1687 struct scatterlist sg
;
1689 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STANDBY
))
1692 addr
= kmemdup(p
, sizeof(*addr
), GFP_KERNEL
);
1696 ret
= eth_prepare_mac_addr_change(dev
, addr
);
1700 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
1701 sg_init_one(&sg
, addr
->sa_data
, dev
->addr_len
);
1702 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1703 VIRTIO_NET_CTRL_MAC_ADDR_SET
, &sg
)) {
1704 dev_warn(&vdev
->dev
,
1705 "Failed to set mac address by vq command.\n");
1709 } else if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
) &&
1710 !virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
1713 /* Naturally, this has an atomicity problem. */
1714 for (i
= 0; i
< dev
->addr_len
; i
++)
1715 virtio_cwrite8(vdev
,
1716 offsetof(struct virtio_net_config
, mac
) +
1717 i
, addr
->sa_data
[i
]);
1720 eth_commit_mac_addr_change(dev
, p
);
1728 static void virtnet_stats(struct net_device
*dev
,
1729 struct rtnl_link_stats64
*tot
)
1731 struct virtnet_info
*vi
= netdev_priv(dev
);
1735 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1736 u64 tpackets
, tbytes
, rpackets
, rbytes
, rdrops
;
1737 struct receive_queue
*rq
= &vi
->rq
[i
];
1738 struct send_queue
*sq
= &vi
->sq
[i
];
1741 start
= u64_stats_fetch_begin_irq(&sq
->stats
.syncp
);
1742 tpackets
= sq
->stats
.packets
;
1743 tbytes
= sq
->stats
.bytes
;
1744 } while (u64_stats_fetch_retry_irq(&sq
->stats
.syncp
, start
));
1747 start
= u64_stats_fetch_begin_irq(&rq
->stats
.syncp
);
1748 rpackets
= rq
->stats
.packets
;
1749 rbytes
= rq
->stats
.bytes
;
1750 rdrops
= rq
->stats
.drops
;
1751 } while (u64_stats_fetch_retry_irq(&rq
->stats
.syncp
, start
));
1753 tot
->rx_packets
+= rpackets
;
1754 tot
->tx_packets
+= tpackets
;
1755 tot
->rx_bytes
+= rbytes
;
1756 tot
->tx_bytes
+= tbytes
;
1757 tot
->rx_dropped
+= rdrops
;
1760 tot
->tx_dropped
= dev
->stats
.tx_dropped
;
1761 tot
->tx_fifo_errors
= dev
->stats
.tx_fifo_errors
;
1762 tot
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1763 tot
->rx_frame_errors
= dev
->stats
.rx_frame_errors
;
1766 static void virtnet_ack_link_announce(struct virtnet_info
*vi
)
1769 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_ANNOUNCE
,
1770 VIRTIO_NET_CTRL_ANNOUNCE_ACK
, NULL
))
1771 dev_warn(&vi
->dev
->dev
, "Failed to ack link announce.\n");
1775 static int _virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1777 struct scatterlist sg
;
1778 struct net_device
*dev
= vi
->dev
;
1780 if (!vi
->has_cvq
|| !virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_MQ
))
1783 vi
->ctrl
->mq
.virtqueue_pairs
= cpu_to_virtio16(vi
->vdev
, queue_pairs
);
1784 sg_init_one(&sg
, &vi
->ctrl
->mq
, sizeof(vi
->ctrl
->mq
));
1786 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MQ
,
1787 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
, &sg
)) {
1788 dev_warn(&dev
->dev
, "Fail to set num of queue pairs to %d\n",
1792 vi
->curr_queue_pairs
= queue_pairs
;
1793 /* virtnet_open() will refill when device is going to up. */
1794 if (dev
->flags
& IFF_UP
)
1795 schedule_delayed_work(&vi
->refill
, 0);
1801 static int virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1806 err
= _virtnet_set_queues(vi
, queue_pairs
);
1811 static int virtnet_close(struct net_device
*dev
)
1813 struct virtnet_info
*vi
= netdev_priv(dev
);
1816 /* Make sure refill_work doesn't re-enable napi! */
1817 cancel_delayed_work_sync(&vi
->refill
);
1819 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1820 xdp_rxq_info_unreg(&vi
->rq
[i
].xdp_rxq
);
1821 napi_disable(&vi
->rq
[i
].napi
);
1822 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1828 static void virtnet_set_rx_mode(struct net_device
*dev
)
1830 struct virtnet_info
*vi
= netdev_priv(dev
);
1831 struct scatterlist sg
[2];
1832 struct virtio_net_ctrl_mac
*mac_data
;
1833 struct netdev_hw_addr
*ha
;
1839 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1840 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_RX
))
1843 vi
->ctrl
->promisc
= ((dev
->flags
& IFF_PROMISC
) != 0);
1844 vi
->ctrl
->allmulti
= ((dev
->flags
& IFF_ALLMULTI
) != 0);
1846 sg_init_one(sg
, &vi
->ctrl
->promisc
, sizeof(vi
->ctrl
->promisc
));
1848 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1849 VIRTIO_NET_CTRL_RX_PROMISC
, sg
))
1850 dev_warn(&dev
->dev
, "Failed to %sable promisc mode.\n",
1851 vi
->ctrl
->promisc
? "en" : "dis");
1853 sg_init_one(sg
, &vi
->ctrl
->allmulti
, sizeof(vi
->ctrl
->allmulti
));
1855 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1856 VIRTIO_NET_CTRL_RX_ALLMULTI
, sg
))
1857 dev_warn(&dev
->dev
, "Failed to %sable allmulti mode.\n",
1858 vi
->ctrl
->allmulti
? "en" : "dis");
1860 uc_count
= netdev_uc_count(dev
);
1861 mc_count
= netdev_mc_count(dev
);
1862 /* MAC filter - use one buffer for both lists */
1863 buf
= kzalloc(((uc_count
+ mc_count
) * ETH_ALEN
) +
1864 (2 * sizeof(mac_data
->entries
)), GFP_ATOMIC
);
1869 sg_init_table(sg
, 2);
1871 /* Store the unicast list and count in the front of the buffer */
1872 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, uc_count
);
1874 netdev_for_each_uc_addr(ha
, dev
)
1875 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1877 sg_set_buf(&sg
[0], mac_data
,
1878 sizeof(mac_data
->entries
) + (uc_count
* ETH_ALEN
));
1880 /* multicast list and count fill the end */
1881 mac_data
= (void *)&mac_data
->macs
[uc_count
][0];
1883 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, mc_count
);
1885 netdev_for_each_mc_addr(ha
, dev
)
1886 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1888 sg_set_buf(&sg
[1], mac_data
,
1889 sizeof(mac_data
->entries
) + (mc_count
* ETH_ALEN
));
1891 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1892 VIRTIO_NET_CTRL_MAC_TABLE_SET
, sg
))
1893 dev_warn(&dev
->dev
, "Failed to set MAC filter table.\n");
1898 static int virtnet_vlan_rx_add_vid(struct net_device
*dev
,
1899 __be16 proto
, u16 vid
)
1901 struct virtnet_info
*vi
= netdev_priv(dev
);
1902 struct scatterlist sg
;
1904 vi
->ctrl
->vid
= cpu_to_virtio16(vi
->vdev
, vid
);
1905 sg_init_one(&sg
, &vi
->ctrl
->vid
, sizeof(vi
->ctrl
->vid
));
1907 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1908 VIRTIO_NET_CTRL_VLAN_ADD
, &sg
))
1909 dev_warn(&dev
->dev
, "Failed to add VLAN ID %d.\n", vid
);
1913 static int virtnet_vlan_rx_kill_vid(struct net_device
*dev
,
1914 __be16 proto
, u16 vid
)
1916 struct virtnet_info
*vi
= netdev_priv(dev
);
1917 struct scatterlist sg
;
1919 vi
->ctrl
->vid
= cpu_to_virtio16(vi
->vdev
, vid
);
1920 sg_init_one(&sg
, &vi
->ctrl
->vid
, sizeof(vi
->ctrl
->vid
));
1922 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1923 VIRTIO_NET_CTRL_VLAN_DEL
, &sg
))
1924 dev_warn(&dev
->dev
, "Failed to kill VLAN ID %d.\n", vid
);
1928 static void virtnet_clean_affinity(struct virtnet_info
*vi
)
1932 if (vi
->affinity_hint_set
) {
1933 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1934 virtqueue_set_affinity(vi
->rq
[i
].vq
, NULL
);
1935 virtqueue_set_affinity(vi
->sq
[i
].vq
, NULL
);
1938 vi
->affinity_hint_set
= false;
1942 static void virtnet_set_affinity(struct virtnet_info
*vi
)
1951 if (!zalloc_cpumask_var(&mask
, GFP_KERNEL
)) {
1952 virtnet_clean_affinity(vi
);
1956 num_cpu
= num_online_cpus();
1957 stride
= max_t(int, num_cpu
/ vi
->curr_queue_pairs
, 1);
1958 stragglers
= num_cpu
>= vi
->curr_queue_pairs
?
1959 num_cpu
% vi
->curr_queue_pairs
:
1961 cpu
= cpumask_next(-1, cpu_online_mask
);
1963 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
1964 group_size
= stride
+ (i
< stragglers
? 1 : 0);
1966 for (j
= 0; j
< group_size
; j
++) {
1967 cpumask_set_cpu(cpu
, mask
);
1968 cpu
= cpumask_next_wrap(cpu
, cpu_online_mask
,
1971 virtqueue_set_affinity(vi
->rq
[i
].vq
, mask
);
1972 virtqueue_set_affinity(vi
->sq
[i
].vq
, mask
);
1973 __netif_set_xps_queue(vi
->dev
, cpumask_bits(mask
), i
, false);
1974 cpumask_clear(mask
);
1977 vi
->affinity_hint_set
= true;
1978 free_cpumask_var(mask
);
1981 static int virtnet_cpu_online(unsigned int cpu
, struct hlist_node
*node
)
1983 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1985 virtnet_set_affinity(vi
);
1989 static int virtnet_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
1991 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1993 virtnet_set_affinity(vi
);
1997 static int virtnet_cpu_down_prep(unsigned int cpu
, struct hlist_node
*node
)
1999 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
2002 virtnet_clean_affinity(vi
);
2006 static enum cpuhp_state virtionet_online
;
2008 static int virtnet_cpu_notif_add(struct virtnet_info
*vi
)
2012 ret
= cpuhp_state_add_instance_nocalls(virtionet_online
, &vi
->node
);
2015 ret
= cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
2019 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
2023 static void virtnet_cpu_notif_remove(struct virtnet_info
*vi
)
2025 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
2026 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
2030 static void virtnet_get_ringparam(struct net_device
*dev
,
2031 struct ethtool_ringparam
*ring
)
2033 struct virtnet_info
*vi
= netdev_priv(dev
);
2035 ring
->rx_max_pending
= virtqueue_get_vring_size(vi
->rq
[0].vq
);
2036 ring
->tx_max_pending
= virtqueue_get_vring_size(vi
->sq
[0].vq
);
2037 ring
->rx_pending
= ring
->rx_max_pending
;
2038 ring
->tx_pending
= ring
->tx_max_pending
;
2042 static void virtnet_get_drvinfo(struct net_device
*dev
,
2043 struct ethtool_drvinfo
*info
)
2045 struct virtnet_info
*vi
= netdev_priv(dev
);
2046 struct virtio_device
*vdev
= vi
->vdev
;
2048 strlcpy(info
->driver
, KBUILD_MODNAME
, sizeof(info
->driver
));
2049 strlcpy(info
->version
, VIRTNET_DRIVER_VERSION
, sizeof(info
->version
));
2050 strlcpy(info
->bus_info
, virtio_bus_name(vdev
), sizeof(info
->bus_info
));
2054 /* TODO: Eliminate OOO packets during switching */
2055 static int virtnet_set_channels(struct net_device
*dev
,
2056 struct ethtool_channels
*channels
)
2058 struct virtnet_info
*vi
= netdev_priv(dev
);
2059 u16 queue_pairs
= channels
->combined_count
;
2062 /* We don't support separate rx/tx channels.
2063 * We don't allow setting 'other' channels.
2065 if (channels
->rx_count
|| channels
->tx_count
|| channels
->other_count
)
2068 if (queue_pairs
> vi
->max_queue_pairs
|| queue_pairs
== 0)
2071 /* For now we don't support modifying channels while XDP is loaded
2072 * also when XDP is loaded all RX queues have XDP programs so we only
2073 * need to check a single RX queue.
2075 if (vi
->rq
[0].xdp_prog
)
2079 err
= _virtnet_set_queues(vi
, queue_pairs
);
2081 netif_set_real_num_tx_queues(dev
, queue_pairs
);
2082 netif_set_real_num_rx_queues(dev
, queue_pairs
);
2084 virtnet_set_affinity(vi
);
2091 static void virtnet_get_strings(struct net_device
*dev
, u32 stringset
, u8
*data
)
2093 struct virtnet_info
*vi
= netdev_priv(dev
);
2094 char *p
= (char *)data
;
2097 switch (stringset
) {
2099 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
2100 for (j
= 0; j
< VIRTNET_RQ_STATS_LEN
; j
++) {
2101 snprintf(p
, ETH_GSTRING_LEN
, "rx_queue_%u_%s",
2102 i
, virtnet_rq_stats_desc
[j
].desc
);
2103 p
+= ETH_GSTRING_LEN
;
2107 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
2108 for (j
= 0; j
< VIRTNET_SQ_STATS_LEN
; j
++) {
2109 snprintf(p
, ETH_GSTRING_LEN
, "tx_queue_%u_%s",
2110 i
, virtnet_sq_stats_desc
[j
].desc
);
2111 p
+= ETH_GSTRING_LEN
;
2118 static int virtnet_get_sset_count(struct net_device
*dev
, int sset
)
2120 struct virtnet_info
*vi
= netdev_priv(dev
);
2124 return vi
->curr_queue_pairs
* (VIRTNET_RQ_STATS_LEN
+
2125 VIRTNET_SQ_STATS_LEN
);
2131 static void virtnet_get_ethtool_stats(struct net_device
*dev
,
2132 struct ethtool_stats
*stats
, u64
*data
)
2134 struct virtnet_info
*vi
= netdev_priv(dev
);
2135 unsigned int idx
= 0, start
, i
, j
;
2136 const u8
*stats_base
;
2139 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
2140 struct receive_queue
*rq
= &vi
->rq
[i
];
2142 stats_base
= (u8
*)&rq
->stats
;
2144 start
= u64_stats_fetch_begin_irq(&rq
->stats
.syncp
);
2145 for (j
= 0; j
< VIRTNET_RQ_STATS_LEN
; j
++) {
2146 offset
= virtnet_rq_stats_desc
[j
].offset
;
2147 data
[idx
+ j
] = *(u64
*)(stats_base
+ offset
);
2149 } while (u64_stats_fetch_retry_irq(&rq
->stats
.syncp
, start
));
2150 idx
+= VIRTNET_RQ_STATS_LEN
;
2153 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
2154 struct send_queue
*sq
= &vi
->sq
[i
];
2156 stats_base
= (u8
*)&sq
->stats
;
2158 start
= u64_stats_fetch_begin_irq(&sq
->stats
.syncp
);
2159 for (j
= 0; j
< VIRTNET_SQ_STATS_LEN
; j
++) {
2160 offset
= virtnet_sq_stats_desc
[j
].offset
;
2161 data
[idx
+ j
] = *(u64
*)(stats_base
+ offset
);
2163 } while (u64_stats_fetch_retry_irq(&sq
->stats
.syncp
, start
));
2164 idx
+= VIRTNET_SQ_STATS_LEN
;
2168 static void virtnet_get_channels(struct net_device
*dev
,
2169 struct ethtool_channels
*channels
)
2171 struct virtnet_info
*vi
= netdev_priv(dev
);
2173 channels
->combined_count
= vi
->curr_queue_pairs
;
2174 channels
->max_combined
= vi
->max_queue_pairs
;
2175 channels
->max_other
= 0;
2176 channels
->rx_count
= 0;
2177 channels
->tx_count
= 0;
2178 channels
->other_count
= 0;
2181 static int virtnet_set_link_ksettings(struct net_device
*dev
,
2182 const struct ethtool_link_ksettings
*cmd
)
2184 struct virtnet_info
*vi
= netdev_priv(dev
);
2186 return ethtool_virtdev_set_link_ksettings(dev
, cmd
,
2187 &vi
->speed
, &vi
->duplex
);
2190 static int virtnet_get_link_ksettings(struct net_device
*dev
,
2191 struct ethtool_link_ksettings
*cmd
)
2193 struct virtnet_info
*vi
= netdev_priv(dev
);
2195 cmd
->base
.speed
= vi
->speed
;
2196 cmd
->base
.duplex
= vi
->duplex
;
2197 cmd
->base
.port
= PORT_OTHER
;
2202 static int virtnet_set_coalesce(struct net_device
*dev
,
2203 struct ethtool_coalesce
*ec
)
2205 struct virtnet_info
*vi
= netdev_priv(dev
);
2208 if (ec
->tx_max_coalesced_frames
> 1 ||
2209 ec
->rx_max_coalesced_frames
!= 1)
2212 napi_weight
= ec
->tx_max_coalesced_frames
? NAPI_POLL_WEIGHT
: 0;
2213 if (napi_weight
^ vi
->sq
[0].napi
.weight
) {
2214 if (dev
->flags
& IFF_UP
)
2216 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2217 vi
->sq
[i
].napi
.weight
= napi_weight
;
2223 static int virtnet_get_coalesce(struct net_device
*dev
,
2224 struct ethtool_coalesce
*ec
)
2226 struct ethtool_coalesce ec_default
= {
2227 .cmd
= ETHTOOL_GCOALESCE
,
2228 .rx_max_coalesced_frames
= 1,
2230 struct virtnet_info
*vi
= netdev_priv(dev
);
2232 memcpy(ec
, &ec_default
, sizeof(ec_default
));
2234 if (vi
->sq
[0].napi
.weight
)
2235 ec
->tx_max_coalesced_frames
= 1;
2240 static void virtnet_init_settings(struct net_device
*dev
)
2242 struct virtnet_info
*vi
= netdev_priv(dev
);
2244 vi
->speed
= SPEED_UNKNOWN
;
2245 vi
->duplex
= DUPLEX_UNKNOWN
;
2248 static void virtnet_update_settings(struct virtnet_info
*vi
)
2253 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_SPEED_DUPLEX
))
2256 speed
= virtio_cread32(vi
->vdev
, offsetof(struct virtio_net_config
,
2258 if (ethtool_validate_speed(speed
))
2260 duplex
= virtio_cread8(vi
->vdev
, offsetof(struct virtio_net_config
,
2262 if (ethtool_validate_duplex(duplex
))
2263 vi
->duplex
= duplex
;
2266 static const struct ethtool_ops virtnet_ethtool_ops
= {
2267 .supported_coalesce_params
= ETHTOOL_COALESCE_MAX_FRAMES
,
2268 .get_drvinfo
= virtnet_get_drvinfo
,
2269 .get_link
= ethtool_op_get_link
,
2270 .get_ringparam
= virtnet_get_ringparam
,
2271 .get_strings
= virtnet_get_strings
,
2272 .get_sset_count
= virtnet_get_sset_count
,
2273 .get_ethtool_stats
= virtnet_get_ethtool_stats
,
2274 .set_channels
= virtnet_set_channels
,
2275 .get_channels
= virtnet_get_channels
,
2276 .get_ts_info
= ethtool_op_get_ts_info
,
2277 .get_link_ksettings
= virtnet_get_link_ksettings
,
2278 .set_link_ksettings
= virtnet_set_link_ksettings
,
2279 .set_coalesce
= virtnet_set_coalesce
,
2280 .get_coalesce
= virtnet_get_coalesce
,
2283 static void virtnet_freeze_down(struct virtio_device
*vdev
)
2285 struct virtnet_info
*vi
= vdev
->priv
;
2288 /* Make sure no work handler is accessing the device */
2289 flush_work(&vi
->config_work
);
2291 netif_tx_lock_bh(vi
->dev
);
2292 netif_device_detach(vi
->dev
);
2293 netif_tx_unlock_bh(vi
->dev
);
2294 cancel_delayed_work_sync(&vi
->refill
);
2296 if (netif_running(vi
->dev
)) {
2297 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2298 napi_disable(&vi
->rq
[i
].napi
);
2299 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
2304 static int init_vqs(struct virtnet_info
*vi
);
2306 static int virtnet_restore_up(struct virtio_device
*vdev
)
2308 struct virtnet_info
*vi
= vdev
->priv
;
2315 virtio_device_ready(vdev
);
2317 if (netif_running(vi
->dev
)) {
2318 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
2319 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
2320 schedule_delayed_work(&vi
->refill
, 0);
2322 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2323 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2324 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
2329 netif_tx_lock_bh(vi
->dev
);
2330 netif_device_attach(vi
->dev
);
2331 netif_tx_unlock_bh(vi
->dev
);
2335 static int virtnet_set_guest_offloads(struct virtnet_info
*vi
, u64 offloads
)
2337 struct scatterlist sg
;
2338 vi
->ctrl
->offloads
= cpu_to_virtio64(vi
->vdev
, offloads
);
2340 sg_init_one(&sg
, &vi
->ctrl
->offloads
, sizeof(vi
->ctrl
->offloads
));
2342 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_GUEST_OFFLOADS
,
2343 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET
, &sg
)) {
2344 dev_warn(&vi
->dev
->dev
, "Fail to set guest offload.\n");
2351 static int virtnet_clear_guest_offloads(struct virtnet_info
*vi
)
2355 if (!vi
->guest_offloads
)
2358 return virtnet_set_guest_offloads(vi
, offloads
);
2361 static int virtnet_restore_guest_offloads(struct virtnet_info
*vi
)
2363 u64 offloads
= vi
->guest_offloads
;
2365 if (!vi
->guest_offloads
)
2368 return virtnet_set_guest_offloads(vi
, offloads
);
2371 static int virtnet_xdp_set(struct net_device
*dev
, struct bpf_prog
*prog
,
2372 struct netlink_ext_ack
*extack
)
2374 unsigned long int max_sz
= PAGE_SIZE
- sizeof(struct padded_vnet_hdr
);
2375 struct virtnet_info
*vi
= netdev_priv(dev
);
2376 struct bpf_prog
*old_prog
;
2377 u16 xdp_qp
= 0, curr_qp
;
2380 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
)
2381 && (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2382 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2383 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2384 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_UFO
) ||
2385 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_CSUM
))) {
2386 NL_SET_ERR_MSG_MOD(extack
, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2390 if (vi
->mergeable_rx_bufs
&& !vi
->any_header_sg
) {
2391 NL_SET_ERR_MSG_MOD(extack
, "XDP expects header/data in single page, any_header_sg required");
2395 if (dev
->mtu
> max_sz
) {
2396 NL_SET_ERR_MSG_MOD(extack
, "MTU too large to enable XDP");
2397 netdev_warn(dev
, "XDP requires MTU less than %lu\n", max_sz
);
2401 curr_qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
;
2403 xdp_qp
= nr_cpu_ids
;
2405 /* XDP requires extra queues for XDP_TX */
2406 if (curr_qp
+ xdp_qp
> vi
->max_queue_pairs
) {
2407 NL_SET_ERR_MSG_MOD(extack
, "Too few free TX rings available");
2408 netdev_warn(dev
, "request %i queues but max is %i\n",
2409 curr_qp
+ xdp_qp
, vi
->max_queue_pairs
);
2413 old_prog
= rtnl_dereference(vi
->rq
[0].xdp_prog
);
2414 if (!prog
&& !old_prog
)
2418 bpf_prog_add(prog
, vi
->max_queue_pairs
- 1);
2420 /* Make sure NAPI is not using any XDP TX queues for RX. */
2421 if (netif_running(dev
)) {
2422 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2423 napi_disable(&vi
->rq
[i
].napi
);
2424 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
2429 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2430 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
2432 virtnet_restore_guest_offloads(vi
);
2437 err
= _virtnet_set_queues(vi
, curr_qp
+ xdp_qp
);
2440 netif_set_real_num_rx_queues(dev
, curr_qp
+ xdp_qp
);
2441 vi
->xdp_queue_pairs
= xdp_qp
;
2444 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2445 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
2446 if (i
== 0 && !old_prog
)
2447 virtnet_clear_guest_offloads(vi
);
2451 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2453 bpf_prog_put(old_prog
);
2454 if (netif_running(dev
)) {
2455 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2456 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
2465 virtnet_clear_guest_offloads(vi
);
2466 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2467 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, old_prog
);
2470 if (netif_running(dev
)) {
2471 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2472 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2473 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
2478 bpf_prog_sub(prog
, vi
->max_queue_pairs
- 1);
2482 static u32
virtnet_xdp_query(struct net_device
*dev
)
2484 struct virtnet_info
*vi
= netdev_priv(dev
);
2485 const struct bpf_prog
*xdp_prog
;
2488 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2489 xdp_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2491 return xdp_prog
->aux
->id
;
2496 static int virtnet_xdp(struct net_device
*dev
, struct netdev_bpf
*xdp
)
2498 switch (xdp
->command
) {
2499 case XDP_SETUP_PROG
:
2500 return virtnet_xdp_set(dev
, xdp
->prog
, xdp
->extack
);
2501 case XDP_QUERY_PROG
:
2502 xdp
->prog_id
= virtnet_xdp_query(dev
);
2509 static int virtnet_get_phys_port_name(struct net_device
*dev
, char *buf
,
2512 struct virtnet_info
*vi
= netdev_priv(dev
);
2515 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STANDBY
))
2518 ret
= snprintf(buf
, len
, "sby");
2525 static int virtnet_set_features(struct net_device
*dev
,
2526 netdev_features_t features
)
2528 struct virtnet_info
*vi
= netdev_priv(dev
);
2532 if ((dev
->features
^ features
) & NETIF_F_LRO
) {
2533 if (vi
->xdp_queue_pairs
)
2536 if (features
& NETIF_F_LRO
)
2537 offloads
= vi
->guest_offloads_capable
;
2541 err
= virtnet_set_guest_offloads(vi
, offloads
);
2544 vi
->guest_offloads
= offloads
;
2550 static const struct net_device_ops virtnet_netdev
= {
2551 .ndo_open
= virtnet_open
,
2552 .ndo_stop
= virtnet_close
,
2553 .ndo_start_xmit
= start_xmit
,
2554 .ndo_validate_addr
= eth_validate_addr
,
2555 .ndo_set_mac_address
= virtnet_set_mac_address
,
2556 .ndo_set_rx_mode
= virtnet_set_rx_mode
,
2557 .ndo_get_stats64
= virtnet_stats
,
2558 .ndo_vlan_rx_add_vid
= virtnet_vlan_rx_add_vid
,
2559 .ndo_vlan_rx_kill_vid
= virtnet_vlan_rx_kill_vid
,
2560 .ndo_bpf
= virtnet_xdp
,
2561 .ndo_xdp_xmit
= virtnet_xdp_xmit
,
2562 .ndo_features_check
= passthru_features_check
,
2563 .ndo_get_phys_port_name
= virtnet_get_phys_port_name
,
2564 .ndo_set_features
= virtnet_set_features
,
2567 static void virtnet_config_changed_work(struct work_struct
*work
)
2569 struct virtnet_info
*vi
=
2570 container_of(work
, struct virtnet_info
, config_work
);
2573 if (virtio_cread_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
,
2574 struct virtio_net_config
, status
, &v
) < 0)
2577 if (v
& VIRTIO_NET_S_ANNOUNCE
) {
2578 netdev_notify_peers(vi
->dev
);
2579 virtnet_ack_link_announce(vi
);
2582 /* Ignore unknown (future) status bits */
2583 v
&= VIRTIO_NET_S_LINK_UP
;
2585 if (vi
->status
== v
)
2590 if (vi
->status
& VIRTIO_NET_S_LINK_UP
) {
2591 virtnet_update_settings(vi
);
2592 netif_carrier_on(vi
->dev
);
2593 netif_tx_wake_all_queues(vi
->dev
);
2595 netif_carrier_off(vi
->dev
);
2596 netif_tx_stop_all_queues(vi
->dev
);
2600 static void virtnet_config_changed(struct virtio_device
*vdev
)
2602 struct virtnet_info
*vi
= vdev
->priv
;
2604 schedule_work(&vi
->config_work
);
2607 static void virtnet_free_queues(struct virtnet_info
*vi
)
2611 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2612 napi_hash_del(&vi
->rq
[i
].napi
);
2613 netif_napi_del(&vi
->rq
[i
].napi
);
2614 netif_napi_del(&vi
->sq
[i
].napi
);
2617 /* We called napi_hash_del() before netif_napi_del(),
2618 * we need to respect an RCU grace period before freeing vi->rq
2627 static void _free_receive_bufs(struct virtnet_info
*vi
)
2629 struct bpf_prog
*old_prog
;
2632 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2633 while (vi
->rq
[i
].pages
)
2634 __free_pages(get_a_page(&vi
->rq
[i
], GFP_KERNEL
), 0);
2636 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2637 RCU_INIT_POINTER(vi
->rq
[i
].xdp_prog
, NULL
);
2639 bpf_prog_put(old_prog
);
2643 static void free_receive_bufs(struct virtnet_info
*vi
)
2646 _free_receive_bufs(vi
);
2650 static void free_receive_page_frags(struct virtnet_info
*vi
)
2653 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2654 if (vi
->rq
[i
].alloc_frag
.page
)
2655 put_page(vi
->rq
[i
].alloc_frag
.page
);
2658 static void free_unused_bufs(struct virtnet_info
*vi
)
2663 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2664 struct virtqueue
*vq
= vi
->sq
[i
].vq
;
2665 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2666 if (!is_xdp_frame(buf
))
2669 xdp_return_frame(ptr_to_xdp(buf
));
2673 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2674 struct virtqueue
*vq
= vi
->rq
[i
].vq
;
2676 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2677 if (vi
->mergeable_rx_bufs
) {
2678 put_page(virt_to_head_page(buf
));
2679 } else if (vi
->big_packets
) {
2680 give_pages(&vi
->rq
[i
], buf
);
2682 put_page(virt_to_head_page(buf
));
2688 static void virtnet_del_vqs(struct virtnet_info
*vi
)
2690 struct virtio_device
*vdev
= vi
->vdev
;
2692 virtnet_clean_affinity(vi
);
2694 vdev
->config
->del_vqs(vdev
);
2696 virtnet_free_queues(vi
);
2699 /* How large should a single buffer be so a queue full of these can fit at
2700 * least one full packet?
2701 * Logic below assumes the mergeable buffer header is used.
2703 static unsigned int mergeable_min_buf_len(struct virtnet_info
*vi
, struct virtqueue
*vq
)
2705 const unsigned int hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2706 unsigned int rq_size
= virtqueue_get_vring_size(vq
);
2707 unsigned int packet_len
= vi
->big_packets
? IP_MAX_MTU
: vi
->dev
->max_mtu
;
2708 unsigned int buf_len
= hdr_len
+ ETH_HLEN
+ VLAN_HLEN
+ packet_len
;
2709 unsigned int min_buf_len
= DIV_ROUND_UP(buf_len
, rq_size
);
2711 return max(max(min_buf_len
, hdr_len
) - hdr_len
,
2712 (unsigned int)GOOD_PACKET_LEN
);
2715 static int virtnet_find_vqs(struct virtnet_info
*vi
)
2717 vq_callback_t
**callbacks
;
2718 struct virtqueue
**vqs
;
2724 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2725 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2726 * possible control vq.
2728 total_vqs
= vi
->max_queue_pairs
* 2 +
2729 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
);
2731 /* Allocate space for find_vqs parameters */
2732 vqs
= kcalloc(total_vqs
, sizeof(*vqs
), GFP_KERNEL
);
2735 callbacks
= kmalloc_array(total_vqs
, sizeof(*callbacks
), GFP_KERNEL
);
2738 names
= kmalloc_array(total_vqs
, sizeof(*names
), GFP_KERNEL
);
2741 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
2742 ctx
= kcalloc(total_vqs
, sizeof(*ctx
), GFP_KERNEL
);
2749 /* Parameters for control virtqueue, if any */
2751 callbacks
[total_vqs
- 1] = NULL
;
2752 names
[total_vqs
- 1] = "control";
2755 /* Allocate/initialize parameters for send/receive virtqueues */
2756 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2757 callbacks
[rxq2vq(i
)] = skb_recv_done
;
2758 callbacks
[txq2vq(i
)] = skb_xmit_done
;
2759 sprintf(vi
->rq
[i
].name
, "input.%d", i
);
2760 sprintf(vi
->sq
[i
].name
, "output.%d", i
);
2761 names
[rxq2vq(i
)] = vi
->rq
[i
].name
;
2762 names
[txq2vq(i
)] = vi
->sq
[i
].name
;
2764 ctx
[rxq2vq(i
)] = true;
2767 ret
= vi
->vdev
->config
->find_vqs(vi
->vdev
, total_vqs
, vqs
, callbacks
,
2773 vi
->cvq
= vqs
[total_vqs
- 1];
2774 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VLAN
))
2775 vi
->dev
->features
|= NETIF_F_HW_VLAN_CTAG_FILTER
;
2778 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2779 vi
->rq
[i
].vq
= vqs
[rxq2vq(i
)];
2780 vi
->rq
[i
].min_buf_len
= mergeable_min_buf_len(vi
, vi
->rq
[i
].vq
);
2781 vi
->sq
[i
].vq
= vqs
[txq2vq(i
)];
2784 /* run here: ret == 0. */
2799 static int virtnet_alloc_queues(struct virtnet_info
*vi
)
2803 vi
->ctrl
= kzalloc(sizeof(*vi
->ctrl
), GFP_KERNEL
);
2806 vi
->sq
= kcalloc(vi
->max_queue_pairs
, sizeof(*vi
->sq
), GFP_KERNEL
);
2809 vi
->rq
= kcalloc(vi
->max_queue_pairs
, sizeof(*vi
->rq
), GFP_KERNEL
);
2813 INIT_DELAYED_WORK(&vi
->refill
, refill_work
);
2814 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2815 vi
->rq
[i
].pages
= NULL
;
2816 netif_napi_add(vi
->dev
, &vi
->rq
[i
].napi
, virtnet_poll
,
2818 netif_tx_napi_add(vi
->dev
, &vi
->sq
[i
].napi
, virtnet_poll_tx
,
2819 napi_tx
? napi_weight
: 0);
2821 sg_init_table(vi
->rq
[i
].sg
, ARRAY_SIZE(vi
->rq
[i
].sg
));
2822 ewma_pkt_len_init(&vi
->rq
[i
].mrg_avg_pkt_len
);
2823 sg_init_table(vi
->sq
[i
].sg
, ARRAY_SIZE(vi
->sq
[i
].sg
));
2825 u64_stats_init(&vi
->rq
[i
].stats
.syncp
);
2826 u64_stats_init(&vi
->sq
[i
].stats
.syncp
);
2839 static int init_vqs(struct virtnet_info
*vi
)
2843 /* Allocate send & receive queues */
2844 ret
= virtnet_alloc_queues(vi
);
2848 ret
= virtnet_find_vqs(vi
);
2853 virtnet_set_affinity(vi
);
2859 virtnet_free_queues(vi
);
2865 static ssize_t
mergeable_rx_buffer_size_show(struct netdev_rx_queue
*queue
,
2868 struct virtnet_info
*vi
= netdev_priv(queue
->dev
);
2869 unsigned int queue_index
= get_netdev_rx_queue_index(queue
);
2870 unsigned int headroom
= virtnet_get_headroom(vi
);
2871 unsigned int tailroom
= headroom
? sizeof(struct skb_shared_info
) : 0;
2872 struct ewma_pkt_len
*avg
;
2874 BUG_ON(queue_index
>= vi
->max_queue_pairs
);
2875 avg
= &vi
->rq
[queue_index
].mrg_avg_pkt_len
;
2876 return sprintf(buf
, "%u\n",
2877 get_mergeable_buf_len(&vi
->rq
[queue_index
], avg
,
2878 SKB_DATA_ALIGN(headroom
+ tailroom
)));
2881 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute
=
2882 __ATTR_RO(mergeable_rx_buffer_size
);
2884 static struct attribute
*virtio_net_mrg_rx_attrs
[] = {
2885 &mergeable_rx_buffer_size_attribute
.attr
,
2889 static const struct attribute_group virtio_net_mrg_rx_group
= {
2890 .name
= "virtio_net",
2891 .attrs
= virtio_net_mrg_rx_attrs
2895 static bool virtnet_fail_on_feature(struct virtio_device
*vdev
,
2897 const char *fname
, const char *dname
)
2899 if (!virtio_has_feature(vdev
, fbit
))
2902 dev_err(&vdev
->dev
, "device advertises feature %s but not %s",
2908 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2909 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2911 static bool virtnet_validate_features(struct virtio_device
*vdev
)
2913 if (!virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) &&
2914 (VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_RX
,
2915 "VIRTIO_NET_F_CTRL_VQ") ||
2916 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_VLAN
,
2917 "VIRTIO_NET_F_CTRL_VQ") ||
2918 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_GUEST_ANNOUNCE
,
2919 "VIRTIO_NET_F_CTRL_VQ") ||
2920 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_MQ
, "VIRTIO_NET_F_CTRL_VQ") ||
2921 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
,
2922 "VIRTIO_NET_F_CTRL_VQ"))) {
2929 #define MIN_MTU ETH_MIN_MTU
2930 #define MAX_MTU ETH_MAX_MTU
2932 static int virtnet_validate(struct virtio_device
*vdev
)
2934 if (!vdev
->config
->get
) {
2935 dev_err(&vdev
->dev
, "%s failure: config access disabled\n",
2940 if (!virtnet_validate_features(vdev
))
2943 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2944 int mtu
= virtio_cread16(vdev
,
2945 offsetof(struct virtio_net_config
,
2948 __virtio_clear_bit(vdev
, VIRTIO_NET_F_MTU
);
2954 static int virtnet_probe(struct virtio_device
*vdev
)
2956 int i
, err
= -ENOMEM
;
2957 struct net_device
*dev
;
2958 struct virtnet_info
*vi
;
2959 u16 max_queue_pairs
;
2962 /* Find if host supports multiqueue virtio_net device */
2963 err
= virtio_cread_feature(vdev
, VIRTIO_NET_F_MQ
,
2964 struct virtio_net_config
,
2965 max_virtqueue_pairs
, &max_queue_pairs
);
2967 /* We need at least 2 queue's */
2968 if (err
|| max_queue_pairs
< VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN
||
2969 max_queue_pairs
> VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX
||
2970 !virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2971 max_queue_pairs
= 1;
2973 /* Allocate ourselves a network device with room for our info */
2974 dev
= alloc_etherdev_mq(sizeof(struct virtnet_info
), max_queue_pairs
);
2978 /* Set up network device as normal. */
2979 dev
->priv_flags
|= IFF_UNICAST_FLT
| IFF_LIVE_ADDR_CHANGE
;
2980 dev
->netdev_ops
= &virtnet_netdev
;
2981 dev
->features
= NETIF_F_HIGHDMA
;
2983 dev
->ethtool_ops
= &virtnet_ethtool_ops
;
2984 SET_NETDEV_DEV(dev
, &vdev
->dev
);
2986 /* Do we support "hardware" checksums? */
2987 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CSUM
)) {
2988 /* This opens up the world of extra features. */
2989 dev
->hw_features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2991 dev
->features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2993 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GSO
)) {
2994 dev
->hw_features
|= NETIF_F_TSO
2995 | NETIF_F_TSO_ECN
| NETIF_F_TSO6
;
2997 /* Individual feature bits: what can host handle? */
2998 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO4
))
2999 dev
->hw_features
|= NETIF_F_TSO
;
3000 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO6
))
3001 dev
->hw_features
|= NETIF_F_TSO6
;
3002 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_ECN
))
3003 dev
->hw_features
|= NETIF_F_TSO_ECN
;
3005 dev
->features
|= NETIF_F_GSO_ROBUST
;
3008 dev
->features
|= dev
->hw_features
& NETIF_F_ALL_TSO
;
3009 /* (!csum && gso) case will be fixed by register_netdev() */
3011 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_CSUM
))
3012 dev
->features
|= NETIF_F_RXCSUM
;
3013 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
3014 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
))
3015 dev
->features
|= NETIF_F_LRO
;
3016 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
))
3017 dev
->hw_features
|= NETIF_F_LRO
;
3019 dev
->vlan_features
= dev
->features
;
3021 /* MTU range: 68 - 65535 */
3022 dev
->min_mtu
= MIN_MTU
;
3023 dev
->max_mtu
= MAX_MTU
;
3025 /* Configuration may specify what MAC to use. Otherwise random. */
3026 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
))
3027 virtio_cread_bytes(vdev
,
3028 offsetof(struct virtio_net_config
, mac
),
3029 dev
->dev_addr
, dev
->addr_len
);
3031 eth_hw_addr_random(dev
);
3033 /* Set up our device-specific information */
3034 vi
= netdev_priv(dev
);
3039 INIT_WORK(&vi
->config_work
, virtnet_config_changed_work
);
3041 /* If we can receive ANY GSO packets, we must allocate large ones. */
3042 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
3043 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
3044 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
3045 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_UFO
))
3046 vi
->big_packets
= true;
3048 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
))
3049 vi
->mergeable_rx_bufs
= true;
3051 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
) ||
3052 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
3053 vi
->hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
3055 vi
->hdr_len
= sizeof(struct virtio_net_hdr
);
3057 if (virtio_has_feature(vdev
, VIRTIO_F_ANY_LAYOUT
) ||
3058 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
3059 vi
->any_header_sg
= true;
3061 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
3064 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
3065 mtu
= virtio_cread16(vdev
,
3066 offsetof(struct virtio_net_config
,
3068 if (mtu
< dev
->min_mtu
) {
3069 /* Should never trigger: MTU was previously validated
3070 * in virtnet_validate.
3073 "device MTU appears to have changed it is now %d < %d",
3081 /* TODO: size buffers correctly in this case. */
3082 if (dev
->mtu
> ETH_DATA_LEN
)
3083 vi
->big_packets
= true;
3086 if (vi
->any_header_sg
)
3087 dev
->needed_headroom
= vi
->hdr_len
;
3089 /* Enable multiqueue by default */
3090 if (num_online_cpus() >= max_queue_pairs
)
3091 vi
->curr_queue_pairs
= max_queue_pairs
;
3093 vi
->curr_queue_pairs
= num_online_cpus();
3094 vi
->max_queue_pairs
= max_queue_pairs
;
3096 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3102 if (vi
->mergeable_rx_bufs
)
3103 dev
->sysfs_rx_queue_group
= &virtio_net_mrg_rx_group
;
3105 netif_set_real_num_tx_queues(dev
, vi
->curr_queue_pairs
);
3106 netif_set_real_num_rx_queues(dev
, vi
->curr_queue_pairs
);
3108 virtnet_init_settings(dev
);
3110 if (virtio_has_feature(vdev
, VIRTIO_NET_F_STANDBY
)) {
3111 vi
->failover
= net_failover_create(vi
->dev
);
3112 if (IS_ERR(vi
->failover
)) {
3113 err
= PTR_ERR(vi
->failover
);
3118 err
= register_netdev(dev
);
3120 pr_debug("virtio_net: registering device failed\n");
3124 virtio_device_ready(vdev
);
3126 err
= virtnet_cpu_notif_add(vi
);
3128 pr_debug("virtio_net: registering cpu notifier failed\n");
3129 goto free_unregister_netdev
;
3132 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
3134 /* Assume link up if device can't report link status,
3135 otherwise get link status from config. */
3136 netif_carrier_off(dev
);
3137 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
)) {
3138 schedule_work(&vi
->config_work
);
3140 vi
->status
= VIRTIO_NET_S_LINK_UP
;
3141 virtnet_update_settings(vi
);
3142 netif_carrier_on(dev
);
3145 for (i
= 0; i
< ARRAY_SIZE(guest_offloads
); i
++)
3146 if (virtio_has_feature(vi
->vdev
, guest_offloads
[i
]))
3147 set_bit(guest_offloads
[i
], &vi
->guest_offloads
);
3148 vi
->guest_offloads_capable
= vi
->guest_offloads
;
3150 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3151 dev
->name
, max_queue_pairs
);
3155 free_unregister_netdev
:
3156 vi
->vdev
->config
->reset(vdev
);
3158 unregister_netdev(dev
);
3160 net_failover_destroy(vi
->failover
);
3162 cancel_delayed_work_sync(&vi
->refill
);
3163 free_receive_page_frags(vi
);
3164 virtnet_del_vqs(vi
);
3170 static void remove_vq_common(struct virtnet_info
*vi
)
3172 vi
->vdev
->config
->reset(vi
->vdev
);
3174 /* Free unused buffers in both send and recv, if any. */
3175 free_unused_bufs(vi
);
3177 free_receive_bufs(vi
);
3179 free_receive_page_frags(vi
);
3181 virtnet_del_vqs(vi
);
3184 static void virtnet_remove(struct virtio_device
*vdev
)
3186 struct virtnet_info
*vi
= vdev
->priv
;
3188 virtnet_cpu_notif_remove(vi
);
3190 /* Make sure no work handler is accessing the device. */
3191 flush_work(&vi
->config_work
);
3193 unregister_netdev(vi
->dev
);
3195 net_failover_destroy(vi
->failover
);
3197 remove_vq_common(vi
);
3199 free_netdev(vi
->dev
);
3202 static __maybe_unused
int virtnet_freeze(struct virtio_device
*vdev
)
3204 struct virtnet_info
*vi
= vdev
->priv
;
3206 virtnet_cpu_notif_remove(vi
);
3207 virtnet_freeze_down(vdev
);
3208 remove_vq_common(vi
);
3213 static __maybe_unused
int virtnet_restore(struct virtio_device
*vdev
)
3215 struct virtnet_info
*vi
= vdev
->priv
;
3218 err
= virtnet_restore_up(vdev
);
3221 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
3223 err
= virtnet_cpu_notif_add(vi
);
3230 static struct virtio_device_id id_table
[] = {
3231 { VIRTIO_ID_NET
, VIRTIO_DEV_ANY_ID
},
3235 #define VIRTNET_FEATURES \
3236 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3238 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3239 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3240 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3241 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3242 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3243 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3244 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3245 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3246 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3248 static unsigned int features
[] = {
3252 static unsigned int features_legacy
[] = {
3255 VIRTIO_F_ANY_LAYOUT
,
3258 static struct virtio_driver virtio_net_driver
= {
3259 .feature_table
= features
,
3260 .feature_table_size
= ARRAY_SIZE(features
),
3261 .feature_table_legacy
= features_legacy
,
3262 .feature_table_size_legacy
= ARRAY_SIZE(features_legacy
),
3263 .driver
.name
= KBUILD_MODNAME
,
3264 .driver
.owner
= THIS_MODULE
,
3265 .id_table
= id_table
,
3266 .validate
= virtnet_validate
,
3267 .probe
= virtnet_probe
,
3268 .remove
= virtnet_remove
,
3269 .config_changed
= virtnet_config_changed
,
3270 #ifdef CONFIG_PM_SLEEP
3271 .freeze
= virtnet_freeze
,
3272 .restore
= virtnet_restore
,
3276 static __init
int virtio_net_driver_init(void)
3280 ret
= cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN
, "virtio/net:online",
3282 virtnet_cpu_down_prep
);
3285 virtionet_online
= ret
;
3286 ret
= cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD
, "virtio/net:dead",
3287 NULL
, virtnet_cpu_dead
);
3291 ret
= register_virtio_driver(&virtio_net_driver
);
3296 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
3298 cpuhp_remove_multi_state(virtionet_online
);
3302 module_init(virtio_net_driver_init
);
3304 static __exit
void virtio_net_driver_exit(void)
3306 unregister_virtio_driver(&virtio_net_driver
);
3307 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
3308 cpuhp_remove_multi_state(virtionet_online
);
3310 module_exit(virtio_net_driver_exit
);
3312 MODULE_DEVICE_TABLE(virtio
, id_table
);
3313 MODULE_DESCRIPTION("Virtio network driver");
3314 MODULE_LICENSE("GPL");